Upstage's low-latency decision model that returns typed choices, scores, and probabilities instead of text, for routing and policy checks.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Why One Forward Pass Per Decision Matters for Gating Latency
Technical Specifications
What the model supports
Comparison
Solar Decide Flash or Solar Decide: Which Decision Model to Deploy?
Solar Decide is the direct sibling. Upstage describes Flash as a faster variant of Solar Decide. Both are built on Solar Mini 4 and use the same typed-question System One approach. Both return choices, scores, or yes/no answers with probabilities in a single forward pass, and both offer a 512K context. OpenRouter lists them at the same input rate with free output. The decision is whether Flash's latency tuning outweighs the more established model.
Choose Solar Decide Flash when decisions sit in a user-facing or high-volume path, such as routing, moderation, or policy gating, and consistent response time matters most. Choose Solar Decide when latency is secondary, or when you want the model Upstage's own System One reference documents as available. Neither has published accuracy benchmarks that we could find, so run both on your labeled examples and compare calibration before committing.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Low-latency request routing
Flash is tuned for consistently fast responses in routing, and each decision is one forward pass. It can pick a team, intent, or downstream model from a defined list of options. Probabilities support confidence thresholds, so low-confidence cases can escalate to a stronger model or a person. It is not suited to routing among more than the documented option cap in a single question.
Policy and moderation checks
Yes/no questions return a probability, which fits policy checks such as urgency, sensitive content, or rule violations. Upstage positions the model for policy checks, and a 524,288-token context lets a whole document be the state. Thresholds should come from your labeled data, because no published accuracy data exists. Decisions that block customers or trigger payments should add human review.
Long-document classification
A 524,288-token context window lets a full contract, report, or thread be classified or scored in one request, without chunking. Choice and score questions turn the document into labels and rubric positions with probabilities. This works well for triage and compliance screening. It will not extract or rewrite passages, since the model produces no text.
Korean-language triage
Flash carries Solar Mini 4's Korean-language strength, per Upstage. It is a natural candidate for classifying and scoring Korean support tickets, messages, and documents. Upstage publishes no Korean benchmark for Flash, so validate on your own labeled data. Write option descriptions as short keywords and check results against your own category definitions.